International Journal of Physical Sciences Vol. 8(5), pp.

186-192, 9 February, 2013

Available online at http://www.academicjournals.org/IJPS
DOI: 10.5897/IJPS12.697
ISSN 1992 - 1950 2013 Academic Journals

Full Length Research Paper

Comparative study of four methods for estimating

Weibull parameters for Halabja, Iraq
Salahaddin A. Ahmed
Department of Physics, Faculty of Science and Science Education, School of Science, University of Sulaimani, Iraq.
E-mail: salahaddinahmed@gmail.com.
Accepted 8 February, 2013

The Weibull distribution is the standard function used by the wind energy community to model the wind
speed frequency distribution. In this study, four methods are presented for estimating Weibull
parameters (Shape and Scale), namely, Maximum likelihood method (MLM), Rank regression method
(RRM), Mean-standard deviation method (MSD), and Power density method (PDM). To compare the
methods, a period of 4 years (2001 - 2004) of monthly time series data of Halabja city was considered.
Two distinct analytical methods are studied to determine the parameter estimation accuracy of these
methods; coefficient of determination and root mean square error (RMSE) are used as measurement
tools. The Rank regression and MSDs are recommended to estimate the shape parameter; also the
Rank regression is recommended for use with our time series wind data to estimate the scale
parameter.

Key words: Weibull distribution, parameter estimation, energy pattern factor, accuracy.

INTRODUCTION

Weibull has been recognized as an appropriate model in
reliability studies and life testing problems such as time to
failure or life length of a component or product. Over the
years, estimation of the shape and scale parameters for a
Weibull distribution function has been approached
through Maximum likelihood method (MLM), linear
method, and several versions of regression analysis. In
recent years, Weibull distribution has been one of the
most commonly used, accepted, recommended
distribution to determine wind energy potential and it is
also used as a reference distribution for commercial wind
energy softwares such as Wind Atlas Analysis and
Application Program (WAsP). The two-parameter Weibull
distribution function is commonly used to fit the wind
speed frequency distribution.
The preferred method of estimating the Weibull
parameters was a graphical way using the cumulative
wind speed distribution, plotting it on special Weibull
graph paper. Estimation of the two-parameter Weibull
distribution occurs in many real-life problems. The
Weibull distribution is an important model especially for
reliability and maintainability analysis.
Weibull distribution can be used to model the wind
speed distribution at a particular site and hence, it can
help in wind resource assessment of a site. By
calculating the two parameters (shape and scale) for
Weibull distribution the wind speed frequency curve for a
site can be made (Prasad et al., 2009) and the key to
perform wind turbine and wind farm energy calculation.
Several methods have been proposed to estimate
Weibull parameters (Marks, 2005; Rider, 1961; Kao,
1959; Pang et al., 2001; Pandey et al., 2011; Seguro and
Lambert, 2000; Stevens and Smulders, 1979;
Bhattacharya and Bhattacharjee, 2010). In literature
about wind energy, these methods are compared several
times and in different ways (Akdag and Ali, 2009; Silva et
al., 2004; Yilmaz et al., 2005; Gupta, 1986; Rahman et
al., 1994; Lei, 2008; Kantar and Senoglu, 2007),
however, results and conclusions of the previous studies
are different. Several of fit tests are used in literature. A
method for estimating parameters of mixed distributions
using sample moments has been outlined by Paul (1961)
who considered compound Poisson, binomial, and a
special case of the mixed Weibull distribution. A graphical
method for estimating the mixed Weibull parameters in
life testing of electron tubes is proposed by John (1959).
For these reasons, according to the results of the studies,
it might be concluded that suitability of the method may
 Ahmed 187

vary with the sample data size, sample data distribution, Taken
sample data format and goodness of fit test (Akdag and k 1
Ali, 2009). v c k 1
x , dV x dx (6)
The present work is based on the time series wind data
c k (6)
collected over a period of 4 years (2001 - 2004) (hourly).
The location concerned in this study named Halabja is Equation 5 can be simplified as:
situated in east Sulaimani/North Iraq 35 11' 7" North 1
latitude, 45 58' 42" East longitude and it is at an V c e x x k
dx (7)
elevation of 692 m above sea level. There is no obstacle 0 (7)
around wind speed measuring location, the wind data
recorded from a mechanical cup type anemometer at This is the form of the standard gamma function, which is given by
height of 2 m above the ground level.

In present study, four methods for estimating the
n e dx (8) (8)
x n 1
parameters of the Weibull wind speed distribution are x
0
presented [MLM, Rank regression method (RRM), Mean-
standard deviation method (MSD), and the Power density 1 the average speed can
method (PDM)] by Akdag' and Ali (2009). The aim of this
From Equations 7 and 8, let n 1
k
work was to select a method that gives more accurate be expressed as:
estimation for the Weibull parameters at this location in
order to reduce uncertainties related to the wind energy 1
V c 1 (9)
output calculation from any Wind Energy Conversion k (9)
Systems (WECS).
The standard deviation of wind speed V is given by

WEIBULL DISTRIBUTION
f
2
v v (v )dv (10)
The Weibull distribution is characterized by two parameters, one is 0
(10)
the scale parameter c (m/s) and the other is the shape parameter k
(dimensionless). In Weibull distribution, the variations in wind speed
are characterized by two functions which are the probability density v f (v )dv 2v vf (v )dv v
2 2

function (PDF) and the cumulative distribution function (CDF). The 0 0

PDF, v,k,c indicates the fraction of time (or probability) for which
the wind is at a given speed V. It is given by Bhattacharya and or
Bhattacharjee (2010) and Weisser (2003).

k V
k 1
c
V
k v f (v )dv 2v .v v
2 2
(11)
f v ,k ,c e (1) 0
(11)
cc (1)
Using
Where > 0, and k, c > 0
The CDF of the speed V gives the fraction of the time (or k 1 k 1
probability) that the wind speed is equal or lower than V, thus, the k v 2 k v
v f (v )dv v dv c 2 x
2 2
dv
k

cumulative distribution v,k,cis the integral of the PDF, given by 0 0 c c 0 c c

V k
FV ,k ,c f V dV 1 e c
Equating to
(2)
0 (2)
e
2 x
c x
2 k
dx (12)
The average wind speed can be expressed as: 0 (12)
2
V V f (V )dV (3) And putting n 1
k
, then the following equation can be
0 (3) obtained. Hence, get the standard deviation

k
V 1
k V k 1 c
1
V 2
2
V ( ) e dV (4) c 2 1 c 2 2 1 (13)
c c k k
0 (4) (13)

This can be rearranged as: or

v
k k

v 2 1
V k e
c
dV (5) c 1
2
1 (14)
0 c
(5) k k
(14)
188 Int. J. Phys. Sci.

METHODS FOR ESTIMATING WEIBULL PARAMETERS n

f ' (k ) ik (ln i ) 2
i 1 k
1 n k
2 i
i 1
1 n
n i 1
n
i 1

k ln i 1 ln i ik ln i (22)
(22)
Maximum likelihood method (MLM)

Maximum likelihood technique, with many required features is the The shape parameter k can be estimated using Equations 21 and
most widely used technique among parameter estimation 22 with Equation 20 as:
techniques. The MLM method has many large sample properties
1
that make it attractive for use; it is asymptotically consistent, which n

ik ln i 1 n
means that as the sample size gets larger, the estimate converges k i 1
ln i (23)
to the true values. n k
i n i 1
Let 1,2,3 ..n be a random sample size n drawn from a i 1 (23)
PDF (,) where is an unknown parameter. The likelihood
function of this random sample is the joint density of the n random Once k is determined, c can be estimated using Equation 18 as
variables and is a function of the unknown parameter. Thus, follows:
(Yilmaz et al., 2005; Nilsen, 2011),
1
1 n k
n c ik (24)
L f v i (v i , ) (15) (15) n i 1 (24)
i 1

The maximum likelihood estimator of say is the value of Rank regression method (RRM)

that maximizes L or, equivalent, the logarithm of L . Often but The second estimation technique, we shall discuss is known as the
not always, the MLM of is a solution of d log L 0 . least squares method. This is, in essence, a more formalized
method of the manual probability plotting technique, in that it
d provides a mathematical method for fitting a line to plotted failure
Now, we apply the MLM to estimate the Weibull parameters, k and data points.
c. Consider the Weibull PDF given in Equation 1, then likelihood It is so commonly applied in engineering and mathematics
function will be (Yilmaz et al., 2005; Nilsen, 2011): problems that are often not thought of as an estimation problem.
k 1
k With the help of this method the parameters are estimated with
k
n i (16) regression line equation by cumulative density function. From
L ( 1 , 2 ,........ n , k ,c ) i e c
(16)
i 1 c c Equation 1, the cumulative density function of Weibull distribution
function with two parameters can be written as (Justus et al., 1978):
On taken the logarithms of Equation 16, differentiating with respect k

to k and c in turn, and equating to zero, one can obtain the
estimating equations F ( i ) 1 e c
(25) (25)

ln L n n 1n This function can be arranged as:

ln i ik ln i 0 (17)
k k i 1 c i 1 (17) k

ln L n 1 n k {1 F ( i )} e1 c
(26) (26)
2 i 0 (18)
c k c i 1 (18)
If we take the natural logarithm of Equation 26
In eliminating c between Equations 17 and 18 and simplifying, one
can get

k

ln{1 F ( i )} i (27)
n c (27)
i ln i
k
1 1 n
i 1
n ln i 0 (19)
i k n i 1
k

i 1 (19)
And then retake the natural logarithm of Equation 27, we get the
This may be solved to get the estimate of k. This can be following equation:
accomplished by the use of standard iterative procedures (that is,
Newton-Raphson method), which can be written in the form
ln[ ln{1 F ( i )}] k ln c k ln i (28) (28)
f (x n )
x n 1 x n (20)
f ' (x n ) (20)
Equation 28 represents a direct relationship between (lni) and
Where
ln{1 F ( i )} which should be minimized
n
i ln i
k
1 1 n
f (k ) i 1
n ln i (21) n

k n i 1 ln[ ln(1 F ( )] ln[ ln(1 E (F ( i ))] (29)

k 2
i i
i 1 (21) i 1 (29)

And Parameters of Weibull distribution with two parameters are

 Ahmed 189

estimated by minimizing with Equation 29. The two parameters c And hence the cubic mean wind speed is given as:
and k are intersecting by the following equations:
3
n n n cu c 3 1 (38)
n ln i ln[ ln{1 F ( i )}] ln i ln[ ln{1 F ( i )}]
k (38)
k i 1
n
i 1
n
i 1
(30)
n ln i2 { ln i }2
i 1 i 1 (30) To determine the energy pattern factor (Epf) one can write
Equations 37 and 38 as:
n n
k ln i ln[ ln{1 F ( i )}]
c exp{ i 1 i 1
} (31) 3
1

nk (31) cu3 k
E pf 3 (39)
1
3 1
k (39)
From Equations 30 and 31, k and c can be estimated, respectively.
Equation 39 is known as energy pattern factor (Epf) method.
Weibull parameters can be estimated with solving energy pattern
Mean-standard deviation method (MSD)
factor Equation 39 numerically or approximately by power density
technique using the simple formula as follows:
The Weibull factors k and c can also be estimated from the mean
and standard deviation of wind data, consider the expression for 3.69
average and standard deviation given in Equations 9 and 14, from k 1 (40)
these, one has (Fung et al., 2007; Weisser and Foxon, 2003): (E pf ) 2
(40)

1
2

Once k is determined, c can be estimated using Equation 37.
v
2
k
1 (32)
2
1
1

k (32)
COMPARISON AND ACCURACY OF THE METHODS
Where
Four methods for estimating the parameters of the Weibull wind
1 n
i (33)
speed distribution for wind energy analysis for Halabja city are
presented. The application of each method is demonstrated using a
n i 1 (33)
sample wind speed data set, and a comparison of the accuracy of
each method is also performed with the actual time series data for
And
the our case study (Halabja city). In order to compare the methods,
monthly mean wind data used for Halabja region is obtained from

2
1 n
2 i
n i 1
(34) meteorological automatic station which covers the period of 4 years
(2001 - 2004).
(34)
Two tests were employed to determine the accuracy of the four
v methods given in this article, first is the coefficient of determination
n is the number of wind observation. Once and are R2 of Equation 41 used to how well the regression model describes
calculated for a given data set, then k can be determined by solving the data, and second is root mean square error (RMSE) of Equation
42.
Equation 32 numerically, once k is determined, c is given by
N

X xi
2
i
R 2 1 i 1
(41)
N

X X
2
c (35) i

1
1 i 1 (41)

k (35)
N
RMSE X xi (42)
2
1
In a simpler approach, an acceptable approximation for k N
i 1
i
(42)
is (Akhlaque et al., 2006):
Where N is the total number of intervals, Xi the frequencies of
1.086 observed wind speed data, xi the frequencies distribution value

k v (36) estimated with Weibull distribution, X the mean of Xi values.
(36)

RESULTS AND DISCUSSION

Power density method (PDM)

This is a new method suggested by Akdag and Ali (2009). It is used Once coefficient of determination and RMSEs are
to estimate the two-Weibull parameters, depends on the energy computed the difference methods can be compared in
pattern factor method; it is related to the averaged data of wind accuracy as shown in Tables 1 and 2. Weibull parameters have
speed. This method has simpler formulation, easier implementation been estimated monthly according to the four methods
and also requires less computation. According to the Weibull with the actual time series data for all the years (2001 -
probability distribution, the mean wind speed of Equation 35 can be
written as (Silva et al., 2004; Paula et al., 2012):
2004). Figure 1 shows the histogram of the actual
frequency distribution of diurnal wind speed for all these
1 years with the Weibull function for fitting a wind data
c 1 (37)
k (37) probability distribution. Figures 2 and 3 show the
190 Int. J. Phys. Sci.

Table 1. Monthly estimated Weibull parameters with actual data.

Mean V Actual Data ML method RR method MSD method PD method

Month
(m/s) K C(m/s) K C(m/s) K C(m/s) K C(m/s) K C(m/s)
Jan 1.49 2.3551 1.6115 1.6072 1.6868 2.9191 1.5669 1.3648 1.6291 1.9102 1.6798
Feb 1.86 1.8691 2.0018 1.4568 2.0841 2.2292 1.9420 1.2725 2.0082 1.8971 1.8688
Mar 1.87 2.3688 2.0723 2.0169 2.1260 2.6248 2.0487 2.0558 2.1113 2.8403 2.0985
April 1.97 2.5514 2.1995 2.2075 2.2420 3.0264 2.1463 2.1368 2.2244 2.9989 2.2053
May 2.16 2.6301 2.4000 2.2757 2.4431 3.1939 2.3338 2.3462 2.4370 3.0965 2.4144
Jun 2.45 3.4864 2.7189 3.1190 2.7445 4.4182 2.6468 3.2168 2.7343 3.6739 2.7146
July 2.43 4.0289 2.6602 3.2868 2.7031 4.6771 2.6293 3.7930 2.6868 3.7933 2.6868
Aug 2.35 3.3551 2.6115 3.1590 2.6240 4.4598 2.5417 3.5062 2.6090 3.7152 2.6038
Sep 2.01 3.0376 2.2200 2.5361 2.2657 3.4402 2.1931 2.6195 2.2622 3.3178 2.2398
Oct 1.75 2.4456 1.9556 2.1106 1.9840 3.1626 1.8745 2.0821 1.9758 2.9119 1.9614
Nov 1.79 2.2320 1.9837 1.9703 2.0281 2.5358 1.9473 1.8807 2.0169 2.7973 2.0087
Dec 1.86 2.1564 2.0682 1.8891 2.1184 2.4237 2.0338 1.7736 2.0908 2.7021 2.0915
Mean 1.99 2.709 2.2086 2.3029 2.2542 3.2592 2.1587 2.3373 2.2322 2.9712 2.2145

Table 2. Statistical analysis for all the methods with actual data.

Standard Coefficient of Coefficient of determination Root mean square error

Method Variance
deviation variation K C (m/s) K C (m/s)
Actual data 0.5977 0.7731 0.4094
ML 0.8459 0.9197 0.4605 0.6290 0.4270 0.3235 0.4464
RR 0.4263 0.6529 0.3374 0.6401 0.4335 0.6328 0.1157
MSD 0.8081 0.8989 0.4544 0.6201 0.4232 0.2534 0.3702
PD 0.5251 0.7246 0.3666 0.5307 0.4088 0.3448 0.3089

estimated parameters c and k, respectively versus the Conclusion

months of years, the similarity can be seen among the
methods with the true data almost for all the months for
parameter c, while for parameter k, the divergence of the
methods with the actual data obtained due to the
difference in the estimated values, these also have
appear in the RMSE results as shown in Table 2. The
rank regression and mean standard deviation methods
give satisfactory results for the shape parameter
estimation, while rank regression method give
satisfactory result for the scale parameter estimation.
Graphically, Figure 4 shows the annual mean wind data
of the probability density function for Weibull distributions
using the estimated Weibull parameters by all methods
have been compared with the annual mean wind true
time series data, as a result the power density method is
the most fitted method to estimate the Weibull
parameters in our study case. We can also see that all
methods are similar enough to show that each method
would be sufficient for determining our parameter
estimates.
According to the results, it might be concluded that
suitability of these methods may vary with the sample
such as data size, sample data distribution (months),
sample data format, and of fit tests. When wind data is
available in time series format, according to the R2 and
RMSE tests both the RRM and MSD, respectively are the
recommended methods for estimating the shape
parameter, while for the scale parameter and for the both
tests, the RRM is recommended method to estimate.
Graphically, the curves of the methods show that the best
way to estimate the two Weibull parameters is the PDM.
This fact is also been supported by means of the RMSE
and R2 statistical tests (Table 2). From this comparative
study, it is observed that the values of RMSE and R2 have
magnitudes that are almost similar for all the methods.
ACKNOWLEDGEMENT
The author would like to thank Dr. Samira Mhamad
 Ahmed 191

5555%

4848%

4141%
Percent of observation
Percent of observation

3434%

2727%

2121%

1414%

7 7%

0 0%
-1 0 1 2 3 4 5 6 7 8 9 10
Wind
Wind vvelocity (m/s)
elocity (m/sec) Figure 4. The Weibull probability density function using all
methods with actual wind data.
Figure 1. Show the histogram of the time series distribution of
the actual wind data.

(Department of Statistics) for providing her information to

this paper.

REFERENCES

Akdag SA, Ali D (2009). A new method to estimate Weibull parameters

for wind energy applications. Energy Convers. Manag. 50:1761-1766.
Akhlaque AM, Firoz A, Wasim AM (2006). Assessment of wind power
potential for coastal areas of Pakistan. Turk. J. Phys. 30:127-135.
Bhattacharya P, Bhattacharjee R (2010). A study on Weibull distribution
for estimating the parameters. J. Appl. Quant. Methods 5(2):234-241.
Fung CH, Yim HL, Lau KH, Kot SC (2007). Win enegy potential in
Guangdong providence. J. Geophys. Res. 112:1-12.
Gupta BK (1986). Weibull parameters for annual and monthly wind
speed distributions for five locations in India. Sol. Energy 37(6):469-
471.
Justus CG, Hargraves W, Amir M, Denise R (1978). Methods for
estimating wind speed frequency distribution. J. Appl. Meteorol.
17:350-353.
Kantar YM, Senoglu B (2007). Estimating the location and scale
Figure 2. Estimated Weibull parameter C (m/s) versus the parameters of the Weibull distribution: an application from
months. engineering. Proceedings of the 16th IASTED International
Conference, Applied Simulation and Modelling. Palma de Mallorca,
Spain.
Kao JH (1959). A graphical estimation of mixed Weibull parameters in
life testing of electron tubes. Technometrics 1:389-407.
Lei Y (2008). Evaluation of three methods for estimating the Weibull
distribution parameters of Chinese pine (Pinus tabulaeformis). J. For.
Sci. 54(12):566-571.
Marks NB (2005). Estimation of Weibull parameters from common
percentiles. J. Appl. Stat. 32(1):17-24.
Nilsen MA (2011). Parameter estimation for the two-parameter Weibull
distribution. Msc. Thesis. Department of Statistics, Brigham Young
University.
Pandey BN, Nidhi D, Pulastya B (2011). Comparison between Bayesian
and maximum likelihood estimation of scale parameter in Weibull
distribution with known shape under linex loss function. J. Sci. Res.
55:163-172.
Pang WK, Jonathan JF, Marvin DT (2001). Estimation of wind speed
distribution using markov chain Monte Carlo techniques. J. Appl.
Meteorol. 40:1476-1484.
Paula AC, Ricardo CS, Carla FD, Maria EV (2012). Comparison of
seven numerical methods for determining Weibull parameters for
wind energy generation I the northeast region of Brazil. Appl. Energy
89:395-400.
Prasad RD, Bansal RC, Sauturaga M (2009). Wind Modeling based on
Figure 3. Estimated Weibull parameter K versus the months. wind input data conditions using Weibull distribution. Inter. J. Glob.
192 Int. J. Phys. Sci.
Energy 32(3):227-240.
Rahman S, Halawani TO, Husain T (1994). Weibull parameters for wind
speed distribution in Saudi Arabia. Sol. Energy 53(6):473-479.
Rider PR (1961). Estimating the parameters of mixed Poisson, Binomail
and Weibull distribution by the method of moments. Bull. Institute Int.
Destatistque 2:38.
Seguro JV, Lambert TW (2000). Modern estimation of parameters of the
Weibull wind speed distribution for wind energy analysis. J. Wind
Eng. Ind. Aerodynamic 85:75-84.
Silva G, Alexandre P, Daniel F, Everaldo F (2004). On the accuracy of
the Weibull parameters estimators. European Wind Energy
Conference and Exhibition (EWEC). London UK.
Stevens MJ, Smulders PT (1979). The estimation of parameters of the
Weibull wind speed distribution for wind energy utilization purpose.
Wind Eng. 3(2):132-145.
Weisser D, Foxon TJ (2003). Implications of seasonal and diurnal
variations of wind velocity for power output estimation of a turbine: A
case study of grenada. Int. J. Energy Res. 2:1165-1179.
Weisser DA (2003). Wind energy analysis of Grenada, an estimation
using Weibull density function. Renew. Energy 28:1803-1812.
Yilmaz V, Aras H, Celik HE (2005). Statistical analysis of wind speed
data. Eng. & Arch. Fac. Eskisehir Osmangazi University. 98(2):1-10.